Since {vex}/{vex3} are respected on legacy mnemonics when -msse2avx is
in use, {evex} should be respected, too. So far this is the case only
for insns where eGPR-s can come into play. Extend coverage to insns with
only %xmm register and possibly immediate operands.
Legacy encoded SIMD insns are converted to AVX ones in that mode. When
eGPR-s are in use, i.e. with APX, convert to AVX10 insns (where
available; there are quite a few which can't be converted).
Note that LDDQU is represented as VMOVDQU32 (and the prior use of the
sse3 template there needs dropping, to get the order right).
Note further that in a few cases, due to the use of templates, AVX512VL
is used when AVX512F would suffice. Since AVX10 is the main reference,
this shouldn't be too much of a problem.
For the case when NDD and NF are both 0 in evex-promoted format,
we will fully support and test it in another patch.
gas/ChangeLog:
* NEWS: Support Intel APX NF.
* config/tc-i386.c (enum i386_error): Add unsupported_nf.
(struct _i386_insn): Add has_nf.
(is_apx_evex_encoding): Ditto.
(build_apx_evex_prefix): Encode the NF bit.
(md_assemble): Handle unsupported_nf.
(parse_insn): Handle Prefix_NF and report bad for illegal combination.
(can_convert_NDD_to_legacy): Replace i.tm.opcode_modifier.nf with i.has_nf.
(match_template): Support D for APX_F insns and check NF support.
* testsuite/gas/i386/x86-64-apx-evex-promoted-bad.d: Add bad test for NF bit.
* testsuite/gas/i386/x86-64-apx-evex-promoted-bad.s: Ditto.
* testsuite/gas/i386/x86-64-apx-inval.l: Ditto.
* testsuite/gas/i386/x86-64-apx-inval.s: Ditto.
* testsuite/gas/i386/x86-64.exp: Add apx nf tests.
* testsuite/gas/i386/x86-64-apx-nf-intel.d: New test.
* testsuite/gas/i386/x86-64-apx-nf.d: Ditto.
* testsuite/gas/i386/x86-64-apx-nf.s: Ditto.
opcodes/ChangeLog:
* i386-dis-evex.h: Add %NF to the instructions that support APX NF and
add new instruction imul, popcnt, tzcnt and lzcnt to EVEX table.
* i386-dis-evex-reg.h: Ditto.
* i386-dis.c (struct instr_info): Add nf.
(struct dis386): Add "NF" for EVEX.NF.
(get_valid_dis386): Set ins->vex.nf and report bad-nf for illegal case.
(print_insn): Handle ins.vex.nf.
(putop): Handle "%NF".
* i386-opc.h (Prefix_NF): New.
* i386-opc.tbl: Added new entries to support full APX NF instructions.
* i386-mnem.h: Regenerated.
* i386-tbl.h: Regenerated.
Like for WRUSSQ it's not needed here. The legacy insn had gained it in
the course of zapping Rex64, but that attribute wasn't needed here
either. The APX insn then simply gained it by copy-and-paste, I suppose.
With the multitude of new APX templates, it finally becomes desirable to
further remove redundancy by also templatizing basic arithmetic insns.
Continue with the shift-double ones.
While there also drop the APX form with ShiftCount omitted. Other shift
and rotate insns were deliberately left without this form as well. Note
that there's also no testsuite adjustment needed for this, indicating
that the form wasn't tested either.
With the multitude of new APX templates, it finally becomes desirable to
further remove redundancy by also templatizing basic arithmetic insns.
Continue with the "ordinary" shift and rotate ones.
While there also drop the APX form of RCL/RCR with Imm1 omitted. Other
shift insns as well as ROR/ROL were deliberately left without this form
as well. Note that there's also no testsuite adjustment needed for this,
indicating that the form wasn't tested either.
Furthermore since RCL/RCR already had non-NDD APX forms, those end up
being added for the other 6 mnemonics, too.
With the multitude of new APX templates, it finally becomes desirable to
further remove redundancy by also templatizing basic arithmetic insns.
Continue with a the more complex binary (two source) cases.
Note how this adds a missing CheckOperandSize to one of the APX sub
forms.
Furthermore since SBB already had a non-NDD APX form, one ends up
being added for the other 6 mnemonics, too.
With the multitude of new APX templates, it finally becomes desirable to
further remove redundancy by also templatizing basic arithmetic insns.
Continue with a few simple unary (single source) cases.
With the multitude of new APX templates, it finally becomes desirable to
further remove redundancy by also templatizing basic arithmetic insns.
Start with the simplest case, accompanied by a necessary adjustment to
i386-gen (such that template uses can also be at the start of a line).
While there also drop a bogus (meaningless / unreachable) "break" as
well as a unused variable (which I'm surprised compilers didn't warn
about).
While most logic in optimize_encoding() is already covering APX by way
of the earlier NDD->REX2 conversion, there's a remaining set of cases
which wants handling separately.
Next to code using %ymm<N> or %zmm<N> it is more natural to have .cfi_*
directives also reference those, not the corresponding %xmm<N>. Accept
their names as kind of aliases, i.e. resolving to the same numbers.
While extending the respective 64-bit testcase, also add %bnd<N> there
(should have happened right with 633789901c ["x86-64: Dwarf2 register
numbers for %bnd<N>"], sorry), requiring binutils/dwarf.c to be adjusted
accordingly as well.
While various other entries in version 003 of the spec aren't quite as
explicit (due to simply leaving the respective field blank), all three
have a clear IGNORED there. IOW they ought to be emitted with EVEX.W=0
by default (and respect -mevexwig=).
While necessary on the legacy encodings, the EVEX ones don't need it.
Even more so when they're available for 64-bit mode only, when the
legacy encodings have the attribute only for correctly handling things
in 16-bit mode.
Several years ago it was decided that SSE2AVX templates should not be
sensitive to -mvexwig= (upon my suggestion to consistently make all
sensitive as long as they don't require a specific setting of VEX.W).
Adjust the four that still are, switching to use of Vex128 at the same
time.
Interestingly unlike VROUND{P,S}{S,D} and VPERM{F,I}128 they weren't
even present in the x86-64-apx-egpr-inval testcase, hence why I
overlooked that these can actually be encoded, (again) using suitable
AVX512 counterparts.
While there also "modernize" the adjacent AVX/AVX2 entries.
With identical source and destination it can be covered by the NDD-to-
legacy conversion logic as well, even if in this case the original insn
doesn't use an NDD encoding. The size savings are even better here, for
the replacement (BSWAP) not having a ModR/M byte.
VRNDSCALE{P,S}{S,D} is the AVX512 generalization of these AVX insns. As
long as the immediate has the top 4 bits clear, they are equivalent to
the earlier VEX-encoded insns, and hence can be used to permit use of
eGPR-s in the memory operand. Since this is the normal way of using
these insns, also alter the resulting diagnostic to complain about the
immediate, not the eGPR use.
Some x86 instructions affect the stack pointer implicitly. Add a new
operand constraint to reflect this. This will be useful for SCFI
implmentation to ensure its correctness.
Mark all push, pop, call, ret, enter, leave, INT, iret instructions.
opcodes/
* i386-gen.c: Update opcode_modifiers.
* i386-opc.h: Add a new constraint.
* i386-opc.tbl: Update the affected instructions.
* i386-tbl.h: Regenerated.
Rex2 is currently an operand constraint. For the upcoming SCFI
implementation in GAS, we need to identify operations which implicitly
update the stack pointer. An operand constraint enumerator for implicit
stack op seems more appropriate than an attribute. However, two opcodes
currently necessitate both Rex2 and an implicit stack op marker; this
prompts revisiting the current representations a bit.
Make Rex2 a standalone attribute, so that later a new operand constraint
may be added for IMPLICIT_STACK_OP.
ChangeLog:
* gas/config/tc-i386.c (is_apx_rex2_encoding): Update the check.
* opcodes/i386-gen.c: Add a new BITFIELD for Rex2.
* opcodes/i386-opc.h (REX2_REQUIRED): Remove.
* opcodes/i386-opc.tbl: Remove Rex2 operand constraint.
* opcodes/i386-tbl.h: Regenerated.
Adds two new external authors to etc/update-copyright.py to cover
bfd/ax_tls.m4, and adds gprofng to dirs handled automatically, then
updates copyright messages as follows:
1) Update cgen/utils.scm emitted copyrights.
2) Run "etc/update-copyright.py --this-year" with an extra external
author I haven't committed, 'Kalray SA.', to cover gas testsuite
files (which should have their copyright message removed).
3) Build with --enable-maintainer-mode --enable-cgen-maint=yes.
4) Check out */po/*.pot which we don't update frequently.
This patch adds non-ND, non-NF forms of EVEX promotion insn.
EVEX extension of legacy instructions:
All promoted legacy instructions are placed in EVEX map 4, which is
currently reserved.
EVEX extension of EVEX instructions:
All existing EVEX instructions are extended by APX using the extended
EVEX prefix, so that they can access all 32 GPRs.
EVEX extension of VEX instructions:
Promoting a VEX instruction into the EVEX space does not change the map
id, the opcode, or the operand encoding of the VEX instruction.
Note: The promoted versions of MOVBE will be extended to include the “MOVBE
reg1, reg2”.
gas/ChangeLog:
2023-12-28 Lingling Kong <lingling.kong@intel.com>
H.J. Lu <hongjiu.lu@intel.com>
Lili Cui <lili.cui@intel.com>
Lin Hu <lin1.hu@intel.com>
* config/tc-i386.c (struct _i386_insn): Add has_egpr.
(need_evex_encoding): Adjusted for apx.
(cpu_flags_match): Ditto.
(install_template): Handled APX combines.
(is_apx_evex_encoding): Test apx evex encoding.
(build_apx_evex_prefix): Enabe APX evex prefix.
(md_assemble): Handle apx with evex encoding.
(process_suffix): Handle apx map4 prefix.
(check_register): Assign i.vec_encoding for APX evex instructions.
* testsuite/gas/i386/x86-64-evex.d: Adjust test cases.
* testsuite/gas/i386/x86-64.exp: Adjust x86-64-inval-movbe.
opcodes/ChangeLog:
* i386-dis-evex-len.h: Handle EVEX_LEN_0F38F2, EVEX_LEN_0F38F3.
* i386-dis-evex-prefix.h: Handle PREFIX_EVEX_0F38F2_L_0,
PREFIX_EVEX_0F38F3_L_0, PREFIX_EVEX_MAP4_D8,
PREFIX_EVEX_MAP4_DA, PREFIX_EVEX_MAP4_DB,
PREFIX_EVEX_MAP4_DC, PREFIX_EVEX_MAP4_DD,
PREFIX_EVEX_MAP4_DE, PREFIX_EVEX_MAP4_DF,
PREFIX_EVEX_MAP4_F0, PREFIX_EVEX_MAP4_F1,
PREFIX_EVEX_MAP4_F2, PREFIX_EVEX_MAP4_F8.
* i386-dis-evex-reg.h: Handle REG_EVEX_0F38F3_L_0_P_0.
* i386-dis-evex.h: Add EVEX_MAP4_ for legacy insn
promote to apx to use gpr32
* opcodes/i386-dis-evex-x86-64.h: Handle Add X86_64_EVEX_0F90,
X86_64_EVEX_0F92, X86_64_EVEX_0F93, X86_64_EVEX_0F38F2,
X86_64_EVEX_0F38F3, X86_64_EVEX_0F38F5, X86_64_EVEX_0F38F6,
X86_64_EVEX_0F38F7, X86_64_EVEX_0F3AF0, X86_64_EVEX_0F91.
* i386-dis.c
(struct instr_info): Deleted bool r.
(PREFIX_NP_OR_DATA): New.
(NO_PREFIX): New.
(putop): Ditto.
(X86_64_EVEX_FROM_VEX_TABLE): Diito.
(get_valid_dis386): Decode insn erex in extend evex prefix.
Handle EVEX_MAP4
(print_insn): Handle PREFIX_DATA_AND_NP_ONLY.
(print_register): Handle apx instructions decode.
(OP_E_memory): Diito.
(OP_G): Diito.
(OP_XMM): Diito.
(DistinctDest_Fixup): Diito.
* i386-gen.c (process_i386_opcode_modifier): Add EVEXMAP4.
* i386-opc.h (SPACE_EVEXMAP4): Add legacy insn
promote to evex.
* i386-opc.tbl: Handle some legacy and vex insns don't
support gpr32. And add some legacy insn (map2 / 3) promote
to evex.
Since AVX10.1/256 will also allow 64 bit mask register, we will
remove the restriction for size of the mask register in AVX10.
gas/ChangeLog:
* config/tc-i386.c (VSZ128, VSZ256, VSZ512): New.
(VEX_check_encoding): Remove opcode_modifier check for vsz.
* testsuite/gas/i386/avx10-vsz.l: Remove testcases for mask
registers since they are not needed.
* testsuite/gas/i386/avx10-vsz.s: Ditto.
opcodes/ChangeLog:
* i386-gen.c: Remove Vsz.
* i386-opc.h: Ditto.
* i386-opc.tbl: Remove kvsz.
* i386-tbl.h: Regenerated.
Now that ATTSyntax and ATTMnemonic aren't use in combination anymore,
fold them and IntelSyntax into a single, enum-like attribute. Note that
this shrinks i386_opcode_modifier back to 2 32-bit words (albeit that's
not for long, seeing in-flight additions for APX).
As noted in the context of d53e6b98a2 ("x86/Intel: correct disassembly
of fsub*/fdiv*") there's no such thing as Intel syntax without Intel
mnemonics. Enforce this on the assembler side, and disentangle command
line option handling on the disassembler side accordingly.
As a result in the opcode table specifying ATTMnemonic|ATTSyntax becomes
redundant with just ATTMnemonic. Drop the now meaningless ATTSyntax and
remove the then no longer accessible templates.
... as MSR index specifier: It is unreasonable to demand that people
write less readable / understandable code, just because the present
documentation mentions only Reg64. Whether to also adjust the
disassembler is a separate question, perhaps indeed more tightly tied
to what the spec says.
Have i386-gen produce merely the offsets into i386_optab[]. Besides
allowing to shrink the table even on 32-bit builds, this results in
removing a level of indirection from the frequently accessed
current_templates, in return for adding a level of indirection when
looking up mnemonics (commonly happening just once per insn). Plus for
PIE builds of gas it also reduces the number of relocations by about two
thousand. Finally a somewhat ugly static variable can also be eliminated
from i386_displacement().
{disp16} is invalid to use in 64-bit mode, while {disp32} is invalid to
use on pre-386 CPUs. The latter, also affecting other (real) prefixes,
further requires that like for insns we fully check the CPU flags; till
now only Cpu64/CpuNo64 were taken into consideration.
As we have grown more uses of it, it becomes increasingly more desirable
to replace it by a simpler check. Have i386-gen do at build time what so
far was done at runtime: Deal with templates indicating EVEX-encoding by
other than the EVex attribute, and set that to "dynamic" in such cases.
This then allows simplifying a number of other conditionals as well.
Right now the opcode table has entries with ISA restrictions of the form
FEAT1|FEAT2, the meaning of which depends on context and requires
special treatment in tc-i386.c: Sometimes this means "both features
requires", whereas originally it was intended to solely mean "all of
these features required". Split the field, with the original one
regaining its original meaning. The new field now truly means "any of
these". The combination of both fields is still and &&-type check, i.e.
(all of these) && (any of these). In the opcode table more involved
combinations of features then also need expressing this way: "all"
entities first, follow by "any" entities enclosed in parentheses, e.g.
x64&(AVX|AVX512F). If the "all" part is empty, parentheses may not be
added around the "any" part (unless parsing logic was further relaxed).
Note that this way AVX512VL no longer needs as much special treatment,
and hence templates previously using AVX512F|AVX512VL are switched to
just AVX512VL.
Note further that this requires FMA handling as resulting from
da0784f961 ("x86: fold FMA VEX and EVEX templates") to be slightly
re-done: FMA now becomes more similar to AVX and AVX2.
Following the folding of some generic AVX/AVX2 templates with their
AVX512F counterpart ones, do this for FMA ones as well, requiring one
further adjustment to cpu_flags_match().
In anticipation of APX introduce logic to reduce the number of templates
we have now, allowing to limit some the number of ones we then need to
gain.
The fundamental requirements are that
- attributes be compatible, which specifically means VexW needs to be
the same in the templates (which often isn't the case, for VEX
encodings having far more WIG tha, EVEX ones),
- the EVEX form being AVX512F (with or without AVX512VL), not any of its
extensions (the same will then be required for APX - it'll need to be
APX_F).
Note that in check_register() there's now a redundant zmm check. Since
this logic will need revisiting for APX anyway, I'd like to keep it that
way for now. (Similarly a couple of if()-s which could be folded are
kept separate, to reduce code churn when adding APX support.)
Now that CpuLM is used solely in cpu_arch_flags and cpu_arch[] while
Cpu64 is solely used in insn templates, they no longer need to be
treated different from other "ordinary" flags; the only "unusual" one
left if CpuNo64. Fold both, leaving just Cpu64.
Recognize "/<number>" suffixes on both -march=+avx10.1 and the
corresponding .arch directive, setting an upper bound on the vector size
that insns may use. Such a restriction can be reset by setting a new base
architecture, by using a suffix-less form, by disabling AVX10, or by
enabling any other VEX/EVEX-based vector extension.
While for most insns we can suppress their use with too wide operands
via registers becoming unavailable (or in Intel syntax memory operand
size specifiers not being recognized), mask register insns have to have
their minimum required vector size specified in a new attribute. (Of
course this new attribute could also be used on other insns.)
Note that .insn continues to be permitted to emit EVEX{512,256} (and
VEX256 ones) encodings regardless of vector size restrictions in place.
Of course these can't be expressed using zmm (or ymm) operands then,
but need using the EVEX.512.* forms (broadcast forms may be usable right
now, but this may go away so shouldn't be relied upon). This is why no
assertions should be added to build_{e,}vex_prefix().
These probably should have been put in place already anyway, but they're
very much wanted in order to then put AVX10.1 support on top. Note that
to avoid reverse dependencies towards SSE (just like we already do for
AVX and XOP), add_isa_dependencies() needs some further tweaking.
While there also address a related anomaly: Disabling AES but neither
AVX nor VAES (similarly for {,V}PCLMULQDQ) would better keep the 128-bit
VEX-encoded forms available. Note that for this the VAES insns are moved
past the AVX+AES ones, to avoid the property-11 test suddenly failing.
The test really is wrong, but let's not also make things inconsistent:
Without the movement, YMM use would be correctly recorded for the
128-bit forms simply because the first template already matches, as long
as VAES wasn't disabled. Yet it still wouldn't be if only AVX+AES were
enabled. Nor would behavior here then be the same as for VPCLMUL* insns.
Commit 916fae9135 ("Add Size64 to movq/vmovq with Reg64 operand" was
right in adding the attribute to MOVQ, but there was no need to add it
to VMOVQ. (See also the AVX512F form, which doesn't have the attribute
either.)
